DeepResearch Bench leaderboard
- Claude Opus 4.6 leads
- 55.3%
- models measured
- 20
- lead over #2
- +0.4 pts
- median
- 46.7%
Leader on DeepResearch Bench · read October 21
On October 21, 2026, 22 of our catalog's models had a DeepResearch Bench result, and Anthropic Claude Opus 4.6 led with 55.3%, ahead of Anthropic Claude Sonnet 4.6 at 54.9%.
Epoch AI collects its results; its benchmarking hub describes the tasks and links the authors.
Role in the index
DeepResearch Bench is shown as published and does not feed any of our indexes. Read the methodology
| # | Model | Result | Evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.6Anthropic | 55.3% | 57.2 | $25.00 | |
| 2 | Claude Sonnet 4.6Anthropic | 54.9% | 57.3 | $15.00 | |
| 3 | Claude Opus 4.5Anthropic | 54.8% | 44.0 | $25.00 | |
| 4 | GPT-5.5OpenAI | 54.0% | 71.3 | $30.00 | |
| 5 | Claude Sonnet 4.5Anthropic | 52.6% | 39.5 | $15.00 | |
| 6 | Claude Opus 4.8Anthropic | 50.2% | 65.0 | $25.00 | |
| 7 | Gemini 3 Flash PreviewGoogle | 49.8% | 43.7 | $3.00 | |
| 8 | GPT-5OpenAI | 49.6% | 38.8 | $10.00 | |
| 9 | Claude Opus 4.1Anthropic | 48.3% | Not benchmarked | $75.00 | |
| 10 | Gemini 3.1 Pro PreviewGoogle | 47.8% | 32.3 | $12.00 | |
| 11 | Claude Opus 4RetiredAnthropic | 46.8% | Not benchmarked | $75.00 | |
| 12 | Claude Haiku 4.5Anthropic | 45.5% | 35.6 | $5.00 | |
| 13 | o3OpenAI | 45.2% | Not benchmarked | $8.00 | |
| 14 | Claude 3.7 SonnetRetiredAnthropic | 43.6% | Not benchmarked | $15.00 | |
| 15 | Gemini 2.5 Pro Preview 06-05Google | 42.8% | Not benchmarked | $10.00 | |
| 16 | GPT-5.1OpenAI | 42.8% | 40.4 | $10.00 | |
| 17 | Gemini 2.5 ProGoogle | 41.5% | 19.1 | $10.00 | |
| 18 | GPT-5.2OpenAI | 41.1% | 55.9 | $14.00 | |
| 19 | Gemini 3.1 Flash LiteGoogle | 37.3% | Not benchmarked | $1.50 | |
| 20 | GPT-5.4 MiniOpenAI | 36.3% | 35.8 | $4.50 | |
| 21 | GPT-5.4OpenAI | 35.1% | 64.4 | $15.00 | |
| 22 | R1 0528DeepSeek | 35.1% | Not benchmarked | $2.18 |
Only models in our catalog are listed, and each appears once with its best published run. 215 rows were read from the source on October 21, 2026, and rows that match no catalog model are left out.